RISC-V: register __asm__ ("tp") doesn't access TP

[palmer-dabbelt]

This came up as part of a glibc patch review https://inbox.sourceware.org/libc-alpha/87wmifdems.fsf@mid.deneb.enyo.de/. Looks like this register ... __asm__ ("tp") syntax isn't working in clang.

void *
__thread_pointer (void)
{
   register void *__tp __asm__ ("tp");
   return __tp;
}

generates

__thread_pointer():
        ret

GCC generates

__thread_pointer():
        mv      a0,tp
        ret

which is the sort of thing I'd expect.

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: Palmer Dabbelt (palmer-dabbelt)

This came up as part of a glibc patch review <https://inbox.sourceware.org/libc-alpha/87wmifdems.fsf@mid.deneb.enyo.de/>. Looks like this `register ... __asm__ ("tp")` syntax isn't working in clang.

void *
__thread_pointer (void)
{
   register void *__tp __asm__ ("tp");
   return __tp;
}

generates

__thread_pointer():
        ret

GCC generates

__thread_pointer():
        mv      a0,tp
        ret

which is the sort of thing I'd expect.

[topperc]

I don't think local register variables work with anything other than inline assembly in clang. A global register variable should work as long as the register isn't allocatable.

[palmer-dabbelt]

I don't think local register variables work with anything other than inline assembly in clang. A global register variable should work as long as the register isn't allocatable.

Ya, that seems to be the conclusion on the glibc side of things too. LLVM is just silently generating unexpected code, though, so I figured it's still worth filing the bug.

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: Palmer Dabbelt (palmer-dabbelt)

This came up as part of a glibc patch review <https://inbox.sourceware.org/libc-alpha/87wmifdems.fsf@mid.deneb.enyo.de/>. Looks like this `register ... __asm__ ("tp")` syntax isn't working in clang.

void *
__thread_pointer (void)
{
   register void *__tp __asm__ ("tp");
   return __tp;
}

generates

__thread_pointer():
        ret

GCC generates

__thread_pointer():
        mv      a0,tp
        ret

which is the sort of thing I'd expect.